CSE-3313--HW09

This is the coding assignment for Homework 9 for CSE3313 (Introduction to Signal Processing).

Purpose

Learn to extract features from the spectrogram of an audio file.

Process

• In the folder All Songs There are 64 songs provided from song-bagpipe.wav to song-winds.wav, along with testSong.wav.

  • The 64 songs of the form song-*.wav, these are the original songs that we will build a database of signatures from.
  • testSong.wav is a sample test file.

• We are using scipy function spectrogram() to analyze each song. f, t, Sxx = spectrogram(x, fs=fs, nperseg=fs//2) where

  • f is an m-element vector of the frequencies evaluated in the signal.
  • t is an n-element vector of the times at which the signal was evaluated.
  • Sxx is an m × n matrix of the amplitudes of each frequency/time pair.

• For each of the 64 songs that fit the pattern song-*.wav we will store its signature and name. This will be our 'database' of songs. Rather than hardcoding all the 64 names, we are using the glob library to build a list of filenames that match the pattern.

• The signature for each song will be a list of the maximum frequencies within each segment of the spectrogram. There will be one value for each of the time segments. For example, if the results of the spectrogram were to produce the plot in

then the signature would be [20, 30, 0] since the max value in col 0 is 8 (frequency = 20Hz), the max val in col 1 is 7 (freq = 30Hz), and the max val in col 3 is 9 (freq = 0Hz).
So, using song-beatles.wav as an example, the signature of that song is (148.0, 146.0, 72.0, 72.0, 108.0, 176.0, 146.0, 148.0, 148.0, 72.0, 72.0, 110.0, 176.0, 148.0, 148.0, 148.0, 72.0, 72.0, 108.0, 176.0, 148.0, 148.0, 108.0, 112.0, 112.0, 132.0, 134.0, 110.0, 112.0, 112.0, 56.0, 112.0, 132.0, 134.0, 110.0, 102.0, 98.0, 98.0, 48.0, 48.0, 48.0, 48.0, 48.0, 50.0, 38.0, 2476.0, 48.0, 52.0, 48.0, 72.0, 72.0, 110.0, 176.0, 176.0, 148.0, 146.0, 74.0, 146.0, 176.0, 176.0, 146.0, 146.0, 72.0, 72.0, 108.0, 176.0, 146.0, 146.0).

• To test our program we will read a file called testSong.wav which is just a copy of one of the corrupted songs matching the pattern test-.wav*.

• The metric used to compare the signatures of the test song and a song in our database is the vector 1-norm. Given a vector , the vector 1-norm is
  

So we will take the 1-norm of the difference of the two signatures. That is, if the signature of the test song is , and the signature of the song from our database is , then we will produce

• We will use this to compare the signature of the test sample with the signature of each song in the database. The results for the 5 songs closest to the test song will be printed to the screen in ascending order. Each line of output consists of
  • the norm value (to zero decimal places)
  • two spaces
  • the name of the song from the database used in the test.

We will also display the spectrograms of the testSong.wav and the first two best matches, each in its own figure window by using matplotlib. pyplot.specgram(x, Fs=fs)